Scholarlink

<p>This study addresses the critical engineering challenge of data scarcity and class imbalance in pediatric dental diagnostics by introducing a novel approach to enhancing both the generation and classification of dental images. To overcome the limitations of traditional datasets, we propose a biologically inspired metaheuristic loss function that emulates natural regenerative processes, specifically drawing from the adaptive healing mechanisms observed in gum tissue regeneration. This loss function is designed to facilitate precise imputation of missing pixels, preserving structural integrity and improving the realism of generated dental images. To further enhance image fidelity, we integrate an 8-connected adaptive method that maintains spatial consistency between pixels, effectively reducing artifacts and preserving fine anatomical details. Additionally, we introduce multiple identity blocks to enhance the model’s adaptability, enabling it to capture complex patterns and variations within dental pathologies more effectively. The practical value of this framework is demonstrated through its ability to significantly diversify training data, thereby improving the generalizability of deep-learning models for reliable clinical diagnostics. Comprehensive experiments conducted on benchmark datasets demonstrate that our approach leads to a substantial improvement in classification accuracy across multiple deep-learning architectures. Specifically, our model achieved an Inception Score (IS) of 76.38 and a Fréchet Inception Distance (FID) of 176.48, indicating effective mode collapse mitigation. With a Peak Signal-to-Noise Ratio (PSNR) of 83.98 and a Structural Similarity Index (SSIM) of 87.87, the average quality score of 85.93 confirms high fidelity of the generated images. Highlighting the engineering utility of this research, the classification results after augmentation showed Vision Transformers achieving an accuracy of 94.76% (up from 87.87%), with sensitivity at 94.65% and specificity at 94.43%. The proposed method outperforms conventional GAN-based augmentation techniques, proving its effectiveness in generating high-quality images that maintain the clinical relevance necessary for automated pediatric dental diagnosis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel approach to pediatric dental imaging: adaptive 8-connected loss and regenerative techniques for improved GAN-based multiple identity block

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Amira Abdelhafeez Elkhatib,
  • Mostafa Elbaz,
  • Reham S. Soliman,
  • Hanaa Salem Marie

摘要

This study addresses the critical engineering challenge of data scarcity and class imbalance in pediatric dental diagnostics by introducing a novel approach to enhancing both the generation and classification of dental images. To overcome the limitations of traditional datasets, we propose a biologically inspired metaheuristic loss function that emulates natural regenerative processes, specifically drawing from the adaptive healing mechanisms observed in gum tissue regeneration. This loss function is designed to facilitate precise imputation of missing pixels, preserving structural integrity and improving the realism of generated dental images. To further enhance image fidelity, we integrate an 8-connected adaptive method that maintains spatial consistency between pixels, effectively reducing artifacts and preserving fine anatomical details. Additionally, we introduce multiple identity blocks to enhance the model’s adaptability, enabling it to capture complex patterns and variations within dental pathologies more effectively. The practical value of this framework is demonstrated through its ability to significantly diversify training data, thereby improving the generalizability of deep-learning models for reliable clinical diagnostics. Comprehensive experiments conducted on benchmark datasets demonstrate that our approach leads to a substantial improvement in classification accuracy across multiple deep-learning architectures. Specifically, our model achieved an Inception Score (IS) of 76.38 and a Fréchet Inception Distance (FID) of 176.48, indicating effective mode collapse mitigation. With a Peak Signal-to-Noise Ratio (PSNR) of 83.98 and a Structural Similarity Index (SSIM) of 87.87, the average quality score of 85.93 confirms high fidelity of the generated images. Highlighting the engineering utility of this research, the classification results after augmentation showed Vision Transformers achieving an accuracy of 94.76% (up from 87.87%), with sensitivity at 94.65% and specificity at 94.43%. The proposed method outperforms conventional GAN-based augmentation techniques, proving its effectiveness in generating high-quality images that maintain the clinical relevance necessary for automated pediatric dental diagnosis.